Fundamenta Informaticae 53(2), 2002, 185-198

Characterization Results for Time-Varying Codes

F.L. T
iplea

(a

E. Makinen
C. Enea

(b

D. Trinca

(a

(a

a)

Faculty of Computer Science

Al.I.Cuza University of Iasi
6600 Iasi, Romania
E-mail: fltiplea,dragost,[email protected]
b)

Department of Computer and Information Sciences

University of Tampere
P.O. Box 607, FIN-33014
Tampere, Finland
E-mail: [email protected]

Abstract
Time-varying codes associate variable length code words to letters being encoded depending on their positions in the input string. These codes
have been introduced in [8] as a proper extension of L-codes.
This paper is devoted to a further study of time-varying codes. First,
we show that adaptive Huffman encodings are special cases of encodings
by time-varying codes. Then, we focus on three kinds of characterization
results: characterization results based on decompositions over families of
sets of words, a Sch
utzenberger like criterion, and a Sardinas-Patterson
like characterization theorem. All of them extend the corresponding characterization results known for classical variable length codes.

Introduction and Preliminaries

Originated in the Shannons information theory in the 1950s, the theory of codes
has developed in several directions. Among them, the theory of variable length
codes, strongly related to combinatorics on words, automata theory, formal languages, and the theory of semigroups, has produced a number of beautiful results
applicable to various fields. Intuitively, a variable length code is a set of words
such that any product of these words can be uniquely decoded.
1

In this paper we deal with time-varying codes, a special class of variable

length codes. These codes have been introduced in [8] as a proper extension
of L-codes [4]. The connection with gsm-codes and SE-codes has been also
discussed in [8]. With a time-varying code, the code word associated to a letter
depends on the position where the letter occurs in the input string. Adaptive
Huffman encodings are special cases of encodings by time-varying codes (Section
2). In Section 3 we focus on characterization results for time-varying codes.
First, we discuss decompositions over families of sets of words, and characterize
them by power-series. Then, two main characterization results are presented:
a Sch
utzenberger like criterion and a Sardinas-Patterson like characterization
theorem for time-varying codes. All these results generalize the ones for classical
codes.
In the rest of this section we recall a few basic concepts and notations on
formal languages and codes (for further details the reader is referred to [1, 5, 6]).
A B denotes the inclusion of A into B, and |A| is the cardinality of the
set A; the empty set is denoted by . The set of natural numbers is denoted by
N. By N we denote the set N {0}.
An alphabet is any nonempty set. For an alphabet , is the free monoid
generated by under the catenation operation, and is its unity (the empty
word).
As usual, + stands for {}. For a set X + , X + is the set
S
n
1
n+1
= X n X, for all n 1 (for two sets of
n1 X , where X = X and X
words A and B, AB is the set of all words uv, where u A and v B). Given
an word w and a set X of words, by wX (Xw, resp.) we denote the set of all
words wx (xw, resp.), where x X.
Let X be a nonempty subset of + , and w + . A decomposition of w
over X is any sequence of words u1 , . . . , us X such that w = u1 us . A code
over is any nonempty subset C + such that each word w + has at
most one decomposition over C. Alternatively, one can say that C is a code
over if there is an alphabet and a function h : + such that f () = C
: of h defined by h()

and the unique homomorphic extension h

=
and
1 n ) = h(1 ) h(n ),
h(
for all 1 n + , is injective. When a code is given by a function f as
above we say that is encoded by f ().
A prefix code is any code C such that no word in C is a proper prefix of
another word in C.
A useful graphic representation of a finite code C + consists of a tree
with vertices labelled by symbols in such that the code words are exactly the
sequences of labels collected from the root to leaves. For example, the tree in
Figure 1 is the graphic representation of the prefix code {01, 110, 101}, where a
is encoded by 01, b by 101, and c by 110.

Time-Varying Codes

Time-Varying Codes (TV-codes, for short) have been introduced in [8] as a

2

t
T
0
T1

T
t

Tt
L
T
0
1
L1
T
L

TTt
Lt t
a C

C1
0
C

Ct t
b c
Figure 1: Tree representation of the code {01, 101, 110}
proper extension of L-codes [4]. The connection to gsm-codes and SE-codes has
been also discussed in [8]. In this section we recall the concept of a TV-code
and show that adaptive Huffman encodings are special cases of encodings by
TV-codes.
A TV-code over an alphabet is any function h : N + , where
: given by h()

is an alphabet, such that the function h

= and
1 n ) = h(1 , 1) h(n , n),
h(
for all 1 n + , is injective. Hence, an input string w over is encoded,
by a TV-code, letter by letter, taking into consideration the letters positions in
w. That is, if occurs on a position i in w, then the occurrence is encoded
by h(, i).
Let h : N + be a TV-code. We denote by Hi the set Hi =
{h(, i)| }, for all i 1, and by H the family H = (Hi |i 1). The set Hi
will be called the ith section of h. When is at most countable we can view h
as an infinite matrix
\ N
1
2

1
h(1 , 1)
h(2 , 1)

2
h(1 , 2)
h(2 , 2)

3
h(1 , 3)
h(2 , 3)

where 1 , 2 , . . . is an arbitrary total ordering of the elements of .

In order to show that the adaptive Huffman encodings are special cases of
encodings by TV-codes, we recall first the concept of a Huffman code [7].
Let be a finite alphabet. A Huffman code for w.r.t. a probability distribution (p | ) on is any prefix code h : {0, 1}+ such that its average
length is minimised. A tree associated to a Huffman code is usually called a
Huffman tree.
The design of a Huffman encoding for an input w + requires two steps:
determine the frequency of occurrences of each letter in w;
3

 design a Huffman code for w.r.t. the probability distribution

p =

frequency of in w
.
|w|

Then, encode w by this Huffman code.

Because this procedure requires two parsings of the input, it is time-consuming
for large inputs (although the compression rate by such an encoding is optimal).
In practice, an alternative method which requires only one parsing of the input
is used. It is called the adaptive Huffman encoding. The encoding of an input
w by adaptive Huffman is based on the construction of a sequence of Huffman
trees as follows (for details the reader is referred to [7]):
start initially with a Huffman tree A0 associated to the alphabet (each
symbol of has frequency 1);
if An is the current Huffman tree and the current input symbol is (that
is, w = uv and u has been already processed), then output the code of
in An (this code is denoted by code(, An )) and update the tree An
getting a new tree An+1 as follows:
increment the frequency of ;
apply the sibling transformation.
The sibling transformation consists of:
1. compare to its successors in the tree (from left to right and from bottom to top). If the immediate successor has frequency k + 1 or greater,
the nodes are still in sorted order and there is no need to change anything. Otherwise, should be swapped with the last successor which has
frequency k or smaller (except that should not be swapped with its
parent);
2. increment the frequency of (from k to k + 1);
3. if is the root, the loop halts; otherwise, the loop repeats with the parent
of .
The adaptive Huffman encoding of w defines naturally a TV-code by
hw (, i) = code(, Ai1 ),
for all , i 1, and taking Ai = A|w| , for all i > |w|.
It is quite easy to prove that hw is a TV-code. Indeed, let us consider two
distinct words over , uu and u u , where , , u, u , u , and
6= . Then,
w (uu ) = h
w (u)hw (, |u| + 1)h
w (u ) = h
w (u)code(, A|u| )h
w (u )
h

and
w (u u ) = h
w (u)hw ( , |u| + 1)h
w (u ) = h
w (u)code( , A|u| )h
w (u ).
h
Because Ai is a prefix code for any i, it follows that neither code(, A|u| ) is a
prefix of code( , A|u| ) nor code( , A|u| ) is a prefix of code(, A|u| ). Therefore,
w (uu ) 6= h
w (u u ), proving that hw is a TV-code.
h
We have proved the following.
Proposition 2.1 Adaptive Huffman encodings are special cases of encodings
by TV-codes.
Example 2.1 In Figure 2, the sequence of Huffman trees needed to encode the
string dcd over the alphabet {a, b, c, d}, is given. The first Huffman tree A0 is
associated to the alphabet. When the first letter d of the input string is read,
it is encoded by code(A0 , d), and a new Huffman tree A1 is generated. This
procedure is iterated until the last letter of the input string is processed. The
4t
T
0 T 1
T

2t
2Tt
L

0 T 1
0 L1
T

T
t
L

1
1 t 1t
1Tt
a
d
b c

5t
T
0 T 1
T

2t
3Tt
L

0 T 1
0 L1
T

T
t
1
1 Lt 1 t
2Tt
a
c
d
b
(b) A1

(a) A0

7t
T
0 T 1

Tt
3 t
4T
T
d
0 T1

2 t
2T t
c
T
0 T 1
T

Tt 1
1 t
a
b

6t
T
0 T 1
T

2t
4Tt
L

0 T 1
0 L1
T

T
t
L

t
t

1
1 2
2Tt
a
d
b c

(c) A3

(d) A2

Figure 2: Huffman trees

TV-code induced by this adaptive Huffman encoding in given in the table below.
\ N
a
b
c
d

1
00
01
10
11

2
00
01
10
11

3
4
00 110
01 111
10 10
11
0

5
110
111
10
0

Characterization Results

The aim of this section is to present several characterization results for TVcodes. First, we characterize the TV-code property by means of decompositions over families of sets of words, and then, a Sch
utzenberger criterion and
a Sardinas-Patterson characterization theorem are presented. All these results
extend the corresponding characterization results known for classical codes.
In order to avoid trivial but annoying analysis cases, the alphabet is
assumed to be of cardinality at least 2 throughout this section.

3.1

Decompositions over Families of Sets of Words

Let A S
= (Ai |i 1) be a family of subsets of + . Denote by Ai the set
i
A = ji Ai Aj . A decomposition of a word w + over the family A is
any sequence of words u1 , . . . , uk such that w = u1 uk and ui Ai , for all
1 i k.
Definition 3.1 A function h : N + is called regular or injective on
sections if the following property holds true:
(i 1)(, )( 6= h(, i) 6= h( , i)).
h is called regular of base C, where C is a nonempty subset of + , if it is regular
and Hi = C, for all i 1.
When is finite, h is regular iff |Hi | = ||, for all i 1.
Remark 3.1 Any TV-code h must be regular. Indeed, if we assume that h is
a TV-code and there are , such that 6= and h(, i) = h( , i) for
some i 1, then
}

} 6=
| {z
| {z
i times
i times

and h( ) = h( ). Hence, h is not a TV-code; a contradiction.

Directly from definitions we have:

Proposition 3.1 A function h : N + is a TV-code iff it is regular
and any word w + has at most one decomposition over its sections H =
(Hi |i 1).
6

Remark 3.2 Without the regularity requirement, the statement in Proposition

3.1 may be false. For example, the function h given in the table below
\ N
1
2

1 2 3
a c c
a d d

is not regular but any word w + has at most one decomposition over its
1 ) = h(
2 ).
family of sections. Moreover, h is not a TV-code because h(
Corollary 3.1 A regular function h : N + of base C is a TV-code iff
C is a code over .
Proof Let h : N + be a regular function of base C. Then, h is a TVcode iff any word w + has at most one decomposition over H = (Hi |i 1)
iff any word w + has at most one decomposition over C iff C is a code over
. 2
Remark 3.3 (1) A section of a TV-code is not necessarily a code. Indeed,
let us consider the function h given in the table below.
\ N
1
2
3

1 2 3
a c c
ab d d
ba e e

It is easily seen that h is a TV-code, but H1 is not a code.

(2) If all sections of a regular function h : N + are prefix codes,
then h is a TV-code. The proof of this fact is similar to the one in Section
2 showing that adaptive Huffman encodings are special cases of encodings
by TV-codes.
The code property of all sections of h does not alone guarantee the TVcode property of h. Indeed, the function h below is regular and each Hi
1 1 ) = h(
2 2 ).
is a code. However, h is not a TV-code because h(
\ N
1
2

1
a
ab

2 3 4
ba c c
a d d

The unique decomposition property over a family of sets of words can be

characterized by formal power-series (for details on semirings and formal powerseries the reader is referred to [2, 3]).
Let be a set and R a semiring. A (formal) power-series over R with
indeterminates inP is defined as a function f : R. An alternative
notation for f is w f (w)w. Let R[[A]] be the set of all power-series over
R with indeterminates in .

The sum and product of two power-series f, g R[[]] are defined by

(f + g)(w) = f (w) + g(w),
X
(f g)(w) =
f (u)g(v),
uv=w

P
for all w (the symbol uv=w in the right hand side of the second
equation indicates summation in the semiring R, over all factorizations uv of
w, and f (u)g(v) is the product of f (u) and g(v) in R). These operations are
associative and R[[]] under them forms a semiring.
In what follows we work only with power-series over the semiring N of natural
numbers with addition and multiplication. Let be an alphabet and X .
The characteristic power-series of X, denoted X , is defined by X (w) = 1 if
w X, and X (w) = 0, otherwise.
Let A1 , . . . , An be subsets of + , where n 2. The product A1 An
is called unambiguous if any word w A1 An has only one decomposition
w = u1 un with ui Ai , for all 1 i n.
The unambiguity property can be easily characterized by power-series as
follows [1]: the product A1 An is unambiguous iff
A1 An = A1 An .
Let h : N + be a function. The family (H1 Hi |i 1) of
formal power-series is locally finite, that is, the following property holds true
(w )(|{i 1|(H1 Hi )(w) 6= 0}| N).
P
Therefore, a power-series i1 H1 Hi can be defined by
X
(
H1 Hi )(w) =
i1

(H1 Hi )(w),

{i1|(H1 Hi )(w)6=0}

for all w (the right hand side of the equality is a finite sum of natural
numbers).

+
Proposition
P 3.2 Let h : N be a regular function. h is a TV-code
iff H 1 = i1 H1 Hi .

Proof
true:

Assume that h is a TV-code. Then, the following two properties hold

for any i 1, the product H1 Hi is unambiguous;

for any w , if w H 1 then there is exactly one i 1 such that
w H1 Hi .

These two properties lead to the following equivalences:

H 1 (w) = 1

w H 1
w H1 Hi , for exactly one i 1
H1 Hi (w) = 1, for exactly one i 1
(
PH1 Hi )(w) = 1, for exactly one i 1
( i1 H1 Hi )(w) = 1,

for any w .
P
Conversely, assume that H 1 = i1 H1 Hi , but h is not a TV-code.
Then, there is a word w + having at least two distinct decompositions over
H. That is, there are i 1 and j 1 such that w H1 Hi H1 Hj (in the
case i = j we assume that w has two distinct decompositions over H1 Hi ).
Then:
(H1 Hi )(w) 1 and (H1 Hj )(w) 1, if j 6= i, and
(H1 Hi )(w) 2, if j = i.
P
Therefore, ( i1 H1 Hi )(w) 2 > 1 = H 1 (w); a contradiction. 2
Remark 3.4 Let h : N + be a regular function of base C.PBy Corollary 3.1 and Proposition 3.1 we obtain that C is a code iff C + = i1 (C )i .
This is a well-known characterization result for classical codes [3].

3.2

Sch
utzenberger Criterion for TV-Codes

Definition 3.2 A family A = (Ai |i 1) of subsets of + is called catenatively

independent if the following property holds
(i 1)(j > i)(Ai Ai Aj = ).
Now, we can prove the following characterization result for TV-codes.
Theorem 3.1 (Sch
utzenberger criterion for TV-codes)
A function h : N + is a TV-code iff the following properties hold:
(1) h is regular;
(2) H is catenatively independent;
(3) for any w + and i, j 1, if w 6 H i+1 then H1 Hi wH1 Hj =
or wH j+1 H i+1 = .
Proof Let us assume that h is a TV-code. Then, clearly, h is regular and H
is catenatively independent.
Let w + and i, j 1 such that w 6 H i+1 . To derive a contradiction,
suppose that H1 Hi w H1 Hj 6= and wH j+1 H i+1 6= . Let x
H1 Hi and y H j+1 such that xw H1 Hj and wy H i+1 . Then,
9

the word xwy has at least two decompositions over H, one of the form x(wy),
beginning by a decomposition of x H1 Hi , and another one of the form
(xw)y, beginning by a decomposition of xw H1 Hj . The decomposition of x
in the word xw is different than the decomposition of x H1 Hi because w is
not a member of H i+1 . Therefore, xwy has at least two distinct decompositions
over H, contradicting the fact that h is a TV-code.
Conversely, suppose that (1), (2), and (3) hold true but h is not a TV-code.
Then, there are two distinct words 1 n , 1 m + such that
h(1 , 1) h(n , n) = h(1 , 1) h(m , m).
Let i be the least index such that i 6= i . By (1), we obtain h(i , i) 6= h(i , i).
Moreover, either h(i , i) is a proper prefix of h(i , i), or vice versa. Let us
suppose that h(i , i) = h(i , i)w, where w is non-empty. Since H is catenatively
independent, w 6 H i+1 .
The equality
h(1 , 1) h(i , i) = h(1 , 1) h(i , i)w
shows that H1 Hi w H1 Hi 6= . Similarly, the equality
h(i+1 , i + 1) h(n , n) = wh(i+1 , i + 1) h(m , m)
shows that wH i+1 H i+1 6= . We have now a contradiction with (2). 2
Remark 3.5 Let C be a nonempty subset of + . Corollary 3.1 and Theorem
3.1 lead to the following conclusion: C is a code iff the following two properties
hold
1. C is catenatively independent;
2. for any w + , if w 6 C + then C + w C + = or wC + C + = .
This is the Sch
utzenberger criterion for codes [1].

3.3

Sardinas-Patterson Criterion for TV-Codes

We are going now to develope a Sardinas-Patterson characterization theorem

for TV-codes. We shall analyse first all the cases where we can get the same
encoding of two different input strings. The encodings will be viewed as two
strings growing successively by catenating new code words.
Case 1: We start by growing both strings by different code words, as in Figure
3. Define the set H1,1 = {x + |, : h(, 1)x = h( , 1)}. The word
w in Figure 3 is a member of this set. Clearly, if H1,1 H2 6= then h is not a
TV-code because it is not catenatively independent.
Case 2: Starting with the configuration in Figure 3 we catenate a new code
word to the first (second) string. Assume that we have the case in Figure 4.
Define the set H2,1 = {x + | : h(, 2)x H1,1 }. The word w in
10

h(, 1)
w
h( , 1)

Figure 3: Generating two distinct code sequences: case 1

h(, 1)

h( , 2)
w
w
h( , 1)

Figure 4: Generating two distinct code sequences: case 2

Figure 4 is a member of H2,1 . Clearly, if H2,1 H3 6= then H is not a TV-code.
Case 3: Starting with the configuration in Figure 3 we catenate a new code
word to the first (second) string. Assume that we have the case in Figure 5.
Define the set H1,2 = {x + |y H1,1 : yx H2 }. The word w in Figure 5
h( , 2)

h(, 1)

h( , 1)

Figure 5: Generating two distinct code sequences: case 3

is a member of H1,2 . Clearly, if H1,2 H2 6= then H is not a TV-code.
The discussion above leads us to consider, for any k 1, the family of sets
Hk = {Hi,j |i, j k} defined as follows:
Hk,k = {x + |Hk x Hk 6= };
Hi,j = {x + |(i > k Hi xHi1,j 6= ) (j > k Hj1,i xHj 6= )},
for all i, j k, but at least one of them greater than k.
Example 3.1 Let h be the function given in the table below.
\ N
1
2
3

1
a
baa
aba

2
3
4
5
baa ab
ab
ab
b
aab aab aab
bab
a
a
a

Then, the following sets are members of the family H1 :

H1,1 = {ba};
H2,1 = {a}, H1,2 = {a, b};
H3,1 = , H2,2 = {aa, ab}, H1,3 = {b, ab};
11

 H4,1 = , H3,2 = {a, b}, H2,3 = {b}, H1,4 = .

Proposition 3.3 Let h : N + be a function. For any k 1, any

i, j k, and any x Hi,j , there are 1 ik+1 , 1 jk+1

+ such that

1 6= 1 and

h(1 , k) h(ik+1 , i)x = h(1 , k) h(jk+1

, j).

Proof We prove the proposition by induction on i + j, where i, j k.

Induction basis: Let x Hk,k . There are 1 , 1 such that h(1 , k)x =
h(1 , k). Because x is non-empty, 1 and 1 should be distinct.
Induction hypothesis: Assume the statement in proposition true for all p, q k
such that p + q = i + j, where where i, j k.
1
Induction step: Let p, q k such that p + q = i + j + 1, and let x Hp,q
. We
have to consider two cases.
Case 1 Hp x Hp1,q 6= . Let y Hp x Hp1,q . Because y Hp1,q ,

by the induction hypothesis there are 1 pk , 1 qk+1

+ such that

1 6= 1 and

h(1 , k) h(pk , p 1)y = h(1 , k) h(qk+1

, q).

Because y Hp x, there is pk+1 such that y = h(pk+1 , p)x. Combining

the above results we have

h(1 , k) h(pk , p 1)h(pk+1 , p)x = h(1 , k) h(qk+1

, q)

which proves the proposition in this case.

Case 2 Hq1,p x Hq 6= . Let y Hq1,p x Hq . Then, there is y Hq1,p

and
such that y = y x. By the induction hypothesis, there are 1 qk
+

1 pk+1 in such that 1 6= 1 and

h(1 , k) h(qk
, q 1)y = h(1 , k) h(pk+1 , p).

Further, we have

h(1 , k) h(qk
, q 1)y x = h(1 , k) h(pk+1 , p)x.

Because y Hq , there is qk+1

such that y = h(qk+1
, q). Therefore, we
obtain

h(1 , k) h(pk+1 , p)x = h(1 , k) h(qk

, q 1)h(qk+1
, q),

which proves the proposition in this case. 2

Now we can prove the following result.
Theorem 3.2 (Sardinas-Patterson criterion for TV-codes)
A function h : N + is a TV-code iff the following properties hold:
12

(1) h is regular;
(2) (k 1)(i, j k)(Hi,j Hi+1 = ).
Proof Let us assume that h is a TV-code. Clearly, h is regular. In order to
prove (2) consider k 1 and i, j k.
To derive a contradiction, assume that Hi,j apHi+1 6= , and let x Hi,j

Hi+1 . By Proposition 3.3, there are 1 ik+1 and 1 jk+1

in + such

that 1 6= 1 and

h(1 , k) h(ik+1 , i)x = h(1 , k) h(jk+1

, j).

Because x Hi+1 , there is such that x = h(, i + 1). Let be an

arbitrary element. Then,
h(, 1) h(, k 1)h(1 , k) h(ik+1 , i)h(, i + 1) =

h(, 1) h(, k 1)h(1 , k) h(jk+1

, j),

is not injective because 1 6= contradicting the fact that

which shows that h
1
h is a TV-code.
Conversely, suppose that (1) and (2) hold true but h is not a TV-code. There

+ such that
are two distinct words 1 n , 1 m
()

1 n ) = h(
).
h(
1
m

Let k be the least number such that k 6= k . Then, () is equivalent to

()

h(k , k) h(n , n) = h(k , k) h(m

, m).

Without loss of the generality we may assume that there are no k i < n and
k j < m such that
h(k , k) h(i , i) = h(k , k) h(j , j).
From (1) it follows that h(k , k) 6= h(k , k), and from () it follows that
h(k , k) is a proper prefix of h(k , k), or vice versa. Let us assume that h(k , k)
is a prefix of h(k , k), and let x be such that h(k , k)x = h(k , k). Then,
x Hk,k , and x 6 Hk+1 because Kk,k Hk+1 = .
The relation () leads to
( )

h(k+1 , k + 1) h(n , n) = xh(k , k) h(m

, m).

Then, h(k+1 , k + 1) is a proper prefix of x, or vice versa. Combining this with

the theorems hypothesis, there is a word y such that y Hk+1,k Hk+2 , if the
first case holds, or y Hk,k+1 Hk+1 , if the second case holds. Both strings in
( ) can now be simplified to the left by h(k+1 , k + 1) resulting

, m)
h(k+2 , k + 2) h(n , n) = yh(k , k) h(m

13

or

yh(k+2 , k + 2) h(n , n) = h(k , k) h(m

, m).

Continuing this process a finite number of times, we get a word z such that

, m) and z Hm1,n Hm .
h(n , n) = z and z Hn1,m Hn , or z = h(m
Both cases lead to a contradiction. Therefore, h is a TV-code. 2
Example 3.2 Consider the function h in Example 3.1. We have H3,2 H4 6= ,
which shows that h is not a TV-code. In fact, it is easy to see that
h(1 , 1)h(2 , 2)h(1 , 3)h(1 , 4) = h(3 , 1)h(2 , 2).
Remark 3.6 We show that the Sardinas-Patterson characterization theorem
for classical codes is a special case of Theorem 3.2.
Let C + be a non-empty set. Define
C1 = {x + |Cx C 6= };
Ci+1 = {x + |Cx Ci 6= Ci x C 6= }, for all i 1.
The Sardinas-Patterson characterization theorem states that C is a code iff
C Ci = , for all i 1.
Assume now that h : N + is a regular function of base C. Then,
Hk,k = H1,1 = C1 , for all k 1. By induction on i 2, we can prove that
Ci =

i
[

Hij+1,j .

j=1

If we assume that Ci1 =

Ci

Si1

j=1

Hij+1,j , then

= {x + |Cx Ci1 6= Ci1 x C 6= }

= {x + |Cx Hi1,1 6=
Cx Hi2,2 6= H1,i1 x C 6=

Cx H1,i1 6= Hi2,2 x C 6=
Cx Hi1,1 6= }
Si
=
j=1 Hij+1,j

Since h is regular and Hk,k = H1,1 = C1 , for all k 1, the condition

(k 1)(i, j 1)(Hi,j Hi+1 = )
is equivalent to
(i 1)(Ci C = ).
Therefore, under the assumption of regularity, Theorem 3.2 is the SardinasPatterson characterization theorem for codes.

14

Conclusions
Time-varying codes associate variable length code words to letters being encoded depending on their positions in the input string. These codes have been
introduced in [8] as a proper extension of L-codes.
In this paper we have continued the study of time-varying codes. First, we
have shown that adaptive Huffman encodings are special cases of encodings by
time-varying codes. Then, we have provided three kinds of characterization results: characterization results based on decompositions over families of sets of
words, a Sch
utzenberger like criterion, and a Sardinas-Patterson like characterization theorem. All of them extend the corresponding characterization results
known for classical variable length codes.

References
[1] J. Berstel, D. Perrin: Theory of Codes, Academic Press, 1985.
[2] W. Kuich, A. Salomaa. Semirings, Automata, and Languages, SpringerVerlag, 1986.
[3] G. Lallement. Semigroups and Combinatorial Applications, John-Wiley &
Sons, 1979.
[4] H.A. Maurer, A. Salomaa, D. Wood. L-Codes and Number Systems, Theoretical Computer Science 22, 1983, 331346.
[5] G. Rozenberg, A. Salomaa (eds.): Handbook of Formal Languages, vol. 1,
Springer-Verlag, 1997.
[6] A. Salomaa: Jewels of Formal Language Theory, Computer Science Press,
1981.
[7] D. Salomon. Data Compression. The Complete Reference, Springer-Verlag,
1998.
[8] F.L. T
iplea, E. Makinen, C. Enea. SE-Systems, Timing Mechanisms, and
Time-Varying Codes, International Journal of Computer Mathematics (to
appear).

15